Processing of Real and AI-Generated Human Faces: An Exploratory EEG Study

Abstract

The increasing presence and quality of AI-generated human faces requires a clearer understanding of how humans perceive these faces. Recent findings indicate AI-generated human faces are rarely effectively distinguished from biological human faces. In this EEG study, 28 participants viewed images of humans, AI-generated humans, robots, and dolls displaying happy, fearful, and neutral expressions. Event-related potentials associated with face processing (N170), expectancy violation and uncanniness (N400), emotionally driven attention (EPN) and sustained attention (LPP) were examined across conditions. Additionally, a questionnaire regarding perceptions of the faces was administered after EEG recordings. This poster presents analyses, some of which explorative, related only to real human and AI-generated human faces. Results yielded significant differences between real and AI-generated human faces in three early ERPs N170, N400, and EPN, but not for the later LPP. Significant differences emerged for expressiveness in emotionally expressive and neutral human and AI-generated faces. Sense of self and inner consciousness was also significantly different across faces. Perceived intelligence and appeal of faces were non-significant. Together, these mixed results indicate that there may be other cognitive processes involved in the processing of human and AI-generated faces than the knowledge or suspicion of realness. Results also indicate that faces are at early processing stages differentiated.